Image Retrieval Based on Relevance Feedback

1. A method of image retrieval, the method comprising: dynamically selecting one of two types of matrixes to be used to weight, based on relevance feedback, a plurality of feature elements for image retrieval, the selecting being based on both a number of previously retrieved relevant images and a length of a feature vector including the plurality of feature elements; and transforming the feature elements to a higher level feature space if the selected one of the two types of matrixes has sufficient information for transformation, but leaving the feature elements untransformed if the selected one of the two types of matrixes has insufficient information for transformation, wherein the selected one of the two types of matrixes having sufficient information for transformation is decomposed as W=P T ΛP, where P is transforming matrix, and Λ is weighting matrix; whereas the selected one of the two types of matrixes having insufficient information for transformation is simply the weighting matrix Λ.

2. A method as recited in claim 1 , wherein the selecting comprises selecting one of the two types of matrixes based on both a number of previously retrieved potentially relevant images which were identified by a user as being relevant, and the length of the feature vector including the plurality of feature elements.

3. A method as recited in claim 1 , wherein the plurality of feature elements are all elements of the same feature.

4. A method as recited in claim 1 , wherein the selecting comprises using a first type of matrix if the number of retrieved relevant images is less than the length of the feature vector, and otherwise using a second type of matrix.

5. A method as recited in claim 4 , wherein the first type of matrix comprises a diagonal matrix and wherein the second type of matrix comprises a full matrix.

6. A method as recited in claim 1 , wherein the selecting comprises using a first type of matrix if the length of the feature vector exceeds the number of retrieved relevant images by at least a threshold amount, and otherwise using a second type of matrix.

7. A method as recited in claim 6 , wherein the first type of matrix comprises a full matrix and the second type of matrix comprises a diagonal matrix.

8. One or more computer readable storage media including a computer program that is executable by a processor to perform the method recited in claim 1 .

9. The method as recited in claim 1 , wherein the Transforming matrix P is an orthonormal matrix consisting of the eigen vectors of W, and weighting matrix Λ is a diagonal matrix whose diagonal elements are the eigen values of W.

10. One or more computer readable storage media having stored thereon a plurality of instructions that, when executed by one or more processors, causes the one or more processors to perform acts including: receiving an initial image selection; generating a plurality of query vectors by extracting, for each query vector, one of a plurality of low-level features from the initial image selection, wherein the low-level features include: a color moments feature, a wavelet based texture feature, and a water-fill edge feature; selecting a set of potentially relevant images based at least in part on distances between the plurality of query vectors and a plurality of feature vectors corresponding to low-level features of a plurality of images; receiving feedback regarding the relevance of one or more images of the set of potentially relevant images; generating a new plurality of query vectors based at least in part on the feedback; generating a weighting of feature elements based at least in part on the feedback; and selecting a new set of potentially relevant images based at least in part on both the weighting of feature elements and distances between the new plurality of query vectors and the plurality of feature vectors, wherein the weighting and the distances are determined using a transformation matrix dynamically selected from two types of matrixes based on both the user feedback and a number of feature elements of each feature vector, and the feature elements are transformed to a higher level feature space if the selected transformation matrix has sufficient information for transformation, but remain untransformed if the selected transformation matrix has insufficient information for transformation.

11. One or more computer readable storage media having stored thereon a plurality of instructions that, when executed by one or more processors, causes the one or more processors to perform acts including: comparing, for each of a plurality of images, a plurality of feature elements from a query vector to a plurality of feature elements from a feature vector corresponding to the image; identifying a number of potentially relevant images based on the comparing; receiving user feedback regarding relevancy of one or more of the potentially relevant images; re-comparing, for each of the plurality of images, the plurality of feature elements from the query vector to the plurality of feature elements from the feature vector, including using a matrix to compare the feature elements, dynamically selecting a type of matrix to use based on both the user feedback and the number of the plurality of feature elements, and transforming the feature elements to a higher level feature space if the selected matrix has sufficient information for transformation, but leaving the feature elements untransformed if the selected matrix has insufficient information for transformation, wherein the dynamically selecting comprises using a first type of matrix if the number of retrieved relevant images is less than the length of the feature vector, and otherwise using a second type of matrix, wherein the first type of matrix comprises a diagonal matrix and the second type of matrix comprises a full matrix identifying a new set of potentially relevant images based on the re-comparing; and presenting the new set of potentially relevant images to the user.

12. One or more computer readable storage media as recited in claim 11 , wherein the re-comparing comprises dynamically selecting the type of matrix to use based on both a number of the potentially relevant images for which user feedback has been received and the number of the plurality of feature elements.

14. A computer as recited in claim 13 , wherein the feature vector includes a plurality of feature elements and wherein the threshold amount comprises the number of feature elements in the feature vector.

16. A computer as recited in claim 13 , further comprising: repeating the generating, identifying of the feature vector, identifying of the number of training samples, and the determining for each of a plurality of features; and identifying how closely the image and the another image match each other by combining the distances between the query vectors and the feature vectors for the plurality of features.

17. A computer as recited in claim 16 , wherein the identifying how closely the image and the another image match each other comprises calculating a weighted summation of each of the individual distances for each of the plurality of features.